24. Cognitive Information Processing

advertisement
Cognitive Information Processing
24. Cognitive Information Processing
Academic and Research Staff
Prof. W.F. Schreiber, Prof. D.E. Troxel, Dr. C.W. Lynn, G. Chambers, C. Konrad, M.
Mcllrath, Dr. K.P. Wacks, R.H. Walker
Graduate Students
R.P. Bishop, R.R. Buckley, G.J. Bunza, Y-M. Chao, R.S. Damon, A. Garcia, P-Q.
Hoang, W.Hofmann, M. Isnardi, M.M.A. Khan, E.A. Lee, D.S. Levinstone, J. Lofton,
E. Peynard, L. Picard, J. Sara, J-P. Schott, G. Vachon, R.E. Velez
24.1 Picture Coding
Donald E. Troxel, William F.Schreiber
a.Sampling and Reconstruction
Most present-day picture coding is digital; virtually all input and output images are analog. Thus,
the conversion from one form to the other iseverpresent and important to the quality and efficiency of
the overall process. To elucidate the phenomena involved, a systematic study was carried out of a
wide variety of pre-sampling and interpolation filters. It was found that, as suspected, the "ideal"
low-pass filter was far from best, subjectively, since the least rms error criterion is invalid for human
observers. Aliasing resulting from "violation" of the sampling theorem is visually traded off against
sharpness and visibility of the sampling structure which occurs with other filters. Several filters
studied performed substantially better that the ILPF. The best combination of performance and
computational simplicity was found in the "sharpened Gaussian" filter, which has an impulse
response not unlike that of human vision.1
b. Differential Pulse Code Modulation
This most common of image data compression systems has been widely studied. Its performance
depends in marked degree on the characteristics of the signal source, since the presence of very
sharp transitions requires an unfavorable trade-off between slope overload and the granular noise
which occurs in blank areas. Nevertheless, when this factor is carefully taken into account, we have
gotten good results with a form of DPCM in which the companding characteristic is adaptively
adjusted to local image properties. In speech coding effective adaptation is possible based only on
the coded signal. Because of the isotropic character of vision and images, this is not possible - extra
data must be transmitted. Even including this extra data, we have obtained nearly "original" quality at
2.2 bits/pel, non-statistical, and 1.5 bits/pel statistical.2
165
RLE P.R. No. 125
Cognitive Information Processing
c. Two-channel Coding System
For a number of years we have studied a picture transmission system in which the signal is divided
into two channels, a two-dimensional low-pass signal ("lows") which is coarsely sampled and finely
quantized, plus the remainder ("highs") which is finely sampled and coarsely quantized with the aid
of a tapered randomized quanitizer. 3 While this can be thought of as a crude form of transform
coding, the two- or three-channel approach permits tailoring the separate channel coding
parameters to well-known human visual properties. While the performance of the non-adaptive
version of this system is not as good as the adaptive DPCM system mentioned above, it does feature
rather simple implementation and excellent (PCM-like) performance in the presence of channel
4
errors. Under a grant from the Sony Corporation, a real-time hardware system was implemented.
d. Adaptive Two-channel Color Coding System
The above-mentioned system can be substantially improved by adapting the highs quantization to
the local signal properties. Blank-area signal-to-noise ratio of 50 db at about 3 bits/pel is possible in
most cases. The system is extended to color coding by transmitting a three-color lows signal plus an
adaptively-companded achromatic highs signal. The color lows signal is further compressed by
mapping to a perceptually uniform color space, akin to the Munsell system of subjective color
notation. Excellent quality full color images are possible at four bits/pel, compared with 24 for the
uncoded signal. Further compression is possible by statistical (entropy) coding of the two-channel
coder output.5
e. Randomized DPCM
We have also studied the application of pseudorandom (PRN) noise to the DPCM quantizer. Unlike
previous efforts to apply PRN to DPCM, we chose to apply the noise directly to the quantizer inside
the feedback loop. The application of PRN greatly reduces the most prominent artifacts of DPCM and
allows greater freedom in the choice on the nonlinear quantizer so that slope overload can be
reduced.6
References
1. J.N. Ratzel, "The Discrete Representation of Spatially Continuous Images," Ph.D. Thesis,
Department of Electrical Engineering and Computer Science, M.I.T., August 1980.
2. A. Zarembowitch, S.M. Thesis, Department of Electrical Engineering and Computer Science, M.I.T.,
1981.
3. D.E. Troxel et al., "Bandwidth Compression of High Quality Images," presented at the International
Conference on Communication, 31.9.1-31.9.5, June 1980.
4. D.E. Troxel et al., "A Two-Channel Picture Coding System: I - Real Time Implementation," IEEE
Trans. on Comm. COM-29, 12, 1841-1848 (1981).
5. W.F. Schreiber and R.R. Buckley, "A Two-Channel Picture Coding System: II - Adaptive
Companding and Color Coding," IEEE Trans. on Comm. COM-9, 12, 1849-1858 (1981).
RLE P.R. No. 125
166
Cognitive Information Processing
6. D.E. Troxel, "Application of Pseudorandom Noise to DPCM," IEEE Trans. on Comm. COM-29, 12,
1763-1767 (1981).
24.2 Digital Wirephoto 24 System
Associated Press (Grant)
Donald E. Troxel, William F.Schrelber
In previous reports, we have discussed the system we developed for the Associated Press for
transmitting pictures to newspapers. This comprises cost-effective Laserphoto 23 facsimile
transmitters and receivers and a computer-based image processing system (Electronic Darkroom)
which permits automatic transmission and reception of pictures over many different channels,
conversion between different Wirephoto scanning standards, and simultaneous editing (enlargement,
reduction, cropping, combining, filtering, tone scale transformation, caption writing, etc.) of stored
images. In this last year of the project, we have continued to improve the system reliability and
performance. The major project was the development of a device which automatically overprints the
computer file name onto facsimile pictures as they are simultaneously received by the computer and a
local facsimile receiver. This greatly reduces the load on the computer system by eliminating the
need to automatically retransmit ID or identification copies to enable the photographic editors to
easily associate pictures with their computer file names. 1
References
1. R.S. Damon, "An Automatic Name Generator for Laserphoto Machines," S.M. Thesis, Department
of Electrical Engineering and Computer Science, M.I.T., August 1982.
24.3 Graphic Arts Applications
William F.Schreiber, Donald E. Troxel
The vast majority of pictures produced every day are made on printing presses. The printing
industry is one of the largest in the country - five times as large as radio and TV broadcasting, and
three times as large as the semiconductor industry. This industry is undergoing a true revolution in
technology based on electronics and computers. For this and other reasons, our activities in recent
years have been focussed on understanding and improving the complicated chain of processes used
to translate images of natural scenes, or man-made graphics, into printed pages. This orientation
affects our activities in a number of ways:
24
Trademark of the A.P.
167
RLE P.R. No. 125
Cognitive Information Processing
1. Our sponsors generally plan to use the products of our research and development in daily
production. Thus, the systems must be practical, reliable, and cost-effective.
2. Graphic arts images are of very high quality compared to those usually used by computer image
processors. This requires careful attention to certain factors, such as tone reproduction, not always
considered of great importance. High quality also implies very large amounts of data to be processed.
3. The images we deal with are intended for human viewing, and the systems we design are always
operated by people. Human perceptual and operational capabilities are central to our work.
The principal operations which must be performed are the scanning and editing of individual page
components, the selection and arrangement of elements on the page (composition, often erroneously
called pagination), the arrangement of pages on the sheet (imposition), and the control of machines
which make printing plates for letterpress, offset, and gravure printing.
24.4 Automated Engraving of Gravure Printing
Providence Gravure, Inc. (Grant)
William F. Schreiber, Donald E. Troxel
Gravure printing is characterized by high platemaking costs but inexpensive and very stable
operation of the printing press. Thus it is suited primarily to long runs, particularly of color work, of
both high and low quality. Substantial economic benefit would accrue from reducing the cost and
time required to prepare printing cylinders, perhaps even making it practical to extend gravure
printing to certain very significant applications such as daily newspapers.
In the system under development, all photographic steps between the original copy and the
cylinder are eliminated by scanning into a computer system. Pictures are interactively edited, and
then all components for each page are assembled into a single disk file. Cylinders are engraved
directly from computer storage. (Inother forms of printing, four color separation films would ba made
for each page instead.) Imposition and correction for ink and paper are performed in "real" time
simultaneously with engraving.
The differences between this system and other existing pre-press systems include substantially
reduced storage requirements due to data compression, high cost-effectiveness, the elimination of
the need for the operators to have long experience in color printing, a high-speed page composition
system, and the ability to perform multiple tasks at the same time. A significant project completed this
year was the development of an interactive silhouetting system. 1 Considerable improvement in the
representation of line art has been made. 2 The monochrome version of the system is in regular daily
production, while the color system is under development.3
?LE P.R. No. 125
168
Cognitive Information Processing
References
1. L. Picard, "Interactive Picture Segmentation," Ph.D. Thesis, Department of Electrical Engineering
and Computer Science, M.I.T., June 1982.
2. J. Sara, "Low Resolution Multi-Level Representation of Bilevel Images," S.M. Thesis, Department
of Electrical Engineering and Computer Science, M.I.T., April 1982.
3. D.E. Troxel et al., "Automated Engraving of Gravure Cylinders," IEEE Trans. on Systems, Man and
Cybernetics SMC-11, 9, 585-596 (1981).
169
RLE P.R. No. 125
RLE P.R. No. 125
Download